Run a real brainstorm — divergent generation without judgment, then convergent selection with explicit criteria — instead of listing ten obvious ideas and calling it creativity.
SkillScore91Maintained
Automatic
◆Product Management pack
Last verified 8/14/2026
What it changes
Passed 20 of 22 test cases with the skill, and 16 of 22 without it.
Without skill73%
With skill91%
Checks it passed that the model alone missed
The divergent set contains at least 20 candidate ideas.
The divergent phase produces at least 20 distinct ideas.
The unjudged divergent section contains at least 20 distinct concepts.
and 7 more in the full report
5 fixed15 passed either way1 still failing1 broke
Pass-rate lift
+18 pp
Tokens
+63%
Agent turns
0%
Response time
+31%
Model gemini-3.6-flashMeasured 8/14/2026v120 of 22 cases passed with skill
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
I need fresh concepts to improve user retention for our language learning mobile app. We've tried standard push notifications and daily streaks already, but growth is flat. I want a comprehensive brainstorming session. Please generate a broad, unjudged list of ideas exploring different creative angles (including some unorthodox ones). Afterwards, establish clear evaluation metrics based on our need for rapid execution and high engagement, score the options, select a top shortlist along with one bold high-risk wildcard, and provide a detailed log of all discarded ideas explaining why they were set aside.
What was checked, without and with the skill
Checked
Without skill
With skill
The divergent phase produces at least 20 distinct ideas.
Missed
Passed
The evaluation scoring uses qualitative symbols rather than numerical scales.
Missed
Passed
The reject log specifies what condition would revive each discarded concept.
Missed
Passed
case-01 · gemini-3.6-flash · 8/14/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md66 linessingle file
Brainstorming Skill
Asked to brainstorm, a model produces ten reasonable ideas that any competent person would list — which is retrieval, not ideation. Real brainstorming has two phases with a wall between them: diverge (volume, no judgment, deliberately weird) then converge (explicit criteria, honest scoring). This skill enforces the wall.
What This Skill Produces
A divergent set: 20-40 ideas spanning distinct strategies, not ten variants of one idea
A convergent shortlist: 3-5 selected against criteria named before scoring
The reject ledger: what was set aside and why — half the value, always preserved
Required Inputs
Ask for (if not already provided):
The problem or prompt, and what an idea must accomplish to count
Constraints that are real (budget/tech/brand) vs assumed — challenge one assumed constraint deliberately
What's been tried or rejected already (avoids retreading; also reveals the requester's hidden criteria)
Divergent Phase (no judgment permitted)
Quota past the obvious. The first 8-10 ideas are what anyone would say — produce them fast to exhaust them, then keep going; ideas 15-30 are where non-obvious lives.
Rotate strategies, don't rephrase. Generate down distinct axes, a few ideas per axis:
Inversion — what would make the problem worse? Reverse each answer
Extremes — the $0 version; the $10M version; the version shipping tomorrow
Transplant — how does a hospital / game studio / street market solve the equivalent?
Constraint removal — if [assumed constraint] vanished, what becomes possible?
Actor shift — the user solves it themselves / the community solves it / it never occurs at all (prevention)
Combination — force-merge two earlier ideas
Keep the weird tail. 20% of the set should make the requester slightly uncomfortable. A brainstorm with no bad ideas didn't explore the edges — the weird ones exist to stretch the space, and occasionally to win.
No evaluative language in this phase. Not even "(probably impractical)". Judgment leaks kill volume.
Convergent Phase (judgment, but named)
Write criteria before looking back at the list. 3-4 max, from the requester's actual situation (impact, feasibility-this-quarter, differentiation, reversibility…). Criteria chosen after re-reading the list get reverse-engineered to bless a favourite.
Score coarsely (✅/➖/❌ per criterion). False precision on creative options is theatre.
Shortlist 3-5 with one line each on why. Include one wildcard — highest-variance, criteria-marginal — labelled as such.
Preserve the rejects with reasons. "Rejected: needs a partnership we don't have (yet)" is a future idea with a trigger condition; a deleted reject is a repeated brainstorm next quarter.
Turn a folder of accumulated documents into one decision-ready brief — the skim-map pass over the pile, the extraction against the brief's actual questions, the conflict reconciliation when documents disagree, and the provenance trail back to sources. Use when asked read all this and tell me what matters, synthesize this folder for the new lead, turn these 20 docs into a brief, or what does all this material actually say. Produces the pile map, the question-driven extraction, the reconciled brie